From 9bec3992312b8bb3aee71bd3b57d106a0a649479 Mon Sep 17 00:00:00 2001 From: Dave Olson Date: Fri, 1 Jun 2007 13:01:47 -0700 Subject: [PATCH] IB/ipath: Verify host bus bandwidth to chip will not limit performance There have been a number of issues where host bandwidth via HT or PCIe to the InfiniPath chip has been limited in some fashion (BIOS, configuration, etc.), resulting in user confusion. This check gives a clear warning that something is wrong and needs to be resolved. Signed-off-by: Dave Olson Signed-off-by: Roland Dreier --- drivers/infiniband/hw/ipath/ipath_driver.c | 86 ++++++++++++++++++++++ 1 file changed, 86 insertions(+) diff --git a/drivers/infiniband/hw/ipath/ipath_driver.c b/drivers/infiniband/hw/ipath/ipath_driver.c index 6ccba365a24..5248f57fe19 100644 --- a/drivers/infiniband/hw/ipath/ipath_driver.c +++ b/drivers/infiniband/hw/ipath/ipath_driver.c @@ -34,6 +34,7 @@ #include #include #include +#include #include #include #include @@ -280,6 +281,89 @@ void __attribute__((weak)) ipath_disable_wc(struct ipath_devdata *dd) { } +/* + * Perform a PIO buffer bandwidth write test, to verify proper system + * configuration. Even when all the setup calls work, occasionally + * BIOS or other issues can prevent write combining from working, or + * can cause other bandwidth problems to the chip. + * + * This test simply writes the same buffer over and over again, and + * measures close to the peak bandwidth to the chip (not testing + * data bandwidth to the wire). On chips that use an address-based + * trigger to send packets to the wire, this is easy. On chips that + * use a count to trigger, we want to make sure that the packet doesn't + * go out on the wire, or trigger flow control checks. + */ +static void ipath_verify_pioperf(struct ipath_devdata *dd) +{ + u32 pbnum, cnt, lcnt; + u32 __iomem *piobuf; + u32 *addr; + u64 msecs, emsecs; + + piobuf = ipath_getpiobuf(dd, &pbnum); + if (!piobuf) { + dev_info(&dd->pcidev->dev, + "No PIObufs for checking perf, skipping\n"); + return; + } + + /* + * Enough to give us a reasonable test, less than piobuf size, and + * likely multiple of store buffer length. + */ + cnt = 1024; + + addr = vmalloc(cnt); + if (!addr) { + dev_info(&dd->pcidev->dev, + "Couldn't get memory for checking PIO perf," + " skipping\n"); + goto done; + } + + preempt_disable(); /* we want reasonably accurate elapsed time */ + msecs = 1 + jiffies_to_msecs(jiffies); + for (lcnt = 0; lcnt < 10000U; lcnt++) { + /* wait until we cross msec boundary */ + if (jiffies_to_msecs(jiffies) >= msecs) + break; + udelay(1); + } + + writeq(0, piobuf); /* length 0, no dwords actually sent */ + ipath_flush_wc(); + + /* + * this is only roughly accurate, since even with preempt we + * still take interrupts that could take a while. Running for + * >= 5 msec seems to get us "close enough" to accurate values + */ + msecs = jiffies_to_msecs(jiffies); + for (emsecs = lcnt = 0; emsecs <= 5UL; lcnt++) { + __iowrite32_copy(piobuf + 64, addr, cnt >> 2); + emsecs = jiffies_to_msecs(jiffies) - msecs; + } + + /* 1 GiB/sec, slightly over IB SDR line rate */ + if (lcnt < (emsecs * 1024U)) + ipath_dev_err(dd, + "Performance problem: bandwidth to PIO buffers is " + "only %u MiB/sec\n", + lcnt / (u32) emsecs); + else + ipath_dbg("PIO buffer bandwidth %u MiB/sec is OK\n", + lcnt / (u32) emsecs); + + preempt_enable(); + + vfree(addr); + +done: + /* disarm piobuf, so it's available again */ + ipath_disarm_piobufs(dd, pbnum, 1); +} + static int __devinit ipath_init_one(struct pci_dev *pdev, const struct pci_device_id *ent) { @@ -515,6 +599,8 @@ static int __devinit ipath_init_one(struct pci_dev *pdev, ret = 0; } + ipath_verify_pioperf(dd); + ipath_device_create_group(&pdev->dev, dd); ipathfs_add_device(dd); ipath_user_add(dd); -- 2.41.1